vllm.model_executor.layers.quantization.gptq_marlin ¶

class GPTQMarlinConfig(QuantizationConfig):
    """Config class for GPTQ Marlin"""

    # (num_bits, is_sym) -> quant_type
    TYPE_MAP = {
        (4, True): scalar_types.uint4b8,
        (8, True): scalar_types.uint8b128,
    }

    def __init__(
        self,
        weight_bits: int,
        group_size: int,
        desc_act: bool,
        is_sym: bool,
        lm_head_quantized: bool,
        dynamic: dict[str, dict[str, int | bool]],
        full_config: dict[str, Any],
        modules_in_block_to_quantize: list[str] | None = None,
    ) -> None:
        super().__init__()
        if desc_act and group_size == -1:
            # In this case, act_order == True is the same as act_order == False
            # (since we have only one group per output channel)
            desc_act = False

        # GPTQModel use `dynamic` config property to allow per module
        # quantization config so each module can be individually optimized.
        # Format is dict[str, dict] where key is a regex string that can
        # perform both positive ("+:" prefixed) or negative ("-:" prefixed)
        # matching of a module.
        # Default to positive match, override base quant config mode, if no
        # prefix is used. Value is in dict format of field key and override
        # value.
        # Negative matching will skip quantization init for this module
        # entirely:
        # non-quantized inference. More details and quantization examples can be
        # found at: https://github.com/ModelCloud/GPTQModel
        # Example:
        #  # last 1/2 of the layers 10-21 has 8bit vs 4bit for 0-9
        #  # last 1/4 of the layers 16-21 has 8bit and group_size 64
        # dynamic = {
        #  #`.*\.` matches the layers_node prefix
        #  # positive match layer 10-15
        #  r"+:.*\.(?:1[0-5])\..*": {"bits": 8,},
        #  # positive match layer 16-21
        #  r"+:.*\.(?:1[6-9]|20|21)\..*": {"bits": 8, "group_size": 64,},
        #  r"-:.*\.moe\..*": {}, # negative match (skip) all `moe` layers
        # }
        self.dynamic = dynamic

        self.weight_bits = weight_bits
        self.is_sym = is_sym

        self.pack_factor = 32 // weight_bits  # packed into int32
        self.group_size = group_size
        self.desc_act = desc_act
        self.lm_head_quantized = lm_head_quantized
        self.full_config = full_config

        if (weight_bits, is_sym) not in self.TYPE_MAP:
            raise ValueError(
                f"Unsupported quantization config: bits={weight_bits}, sym={is_sym}"
            )

        self.quant_type = self.TYPE_MAP[(weight_bits, is_sym)]

        self.modules_in_block_to_quantize = modules_in_block_to_quantize or []
        # used to identify GPTQ model quantized by autoround
        self.autoround_version = full_config.get("autoround_version", "")

    def __repr__(self) -> str:
        return (
            f"GPTQMarlinConfig(quant_type={self.quant_type}, "
            f"group_size={self.group_size}, "
            f"desc_act={self.desc_act}, "
            f"lm_head_quantized={self.lm_head_quantized}, "
            f"dynamic={self.dynamic}, "
            f"modules_in_block_to_quantize={self.modules_in_block_to_quantize})"
        )

    @classmethod
    def get_name(cls) -> QuantizationMethods:
        return "gptq_marlin"

    @classmethod
    def get_supported_act_dtypes(cls) -> list[torch.dtype]:
        return [torch.half, torch.bfloat16]

    @classmethod
    def get_min_capability(cls) -> int:
        return 80

    @classmethod
    def get_config_filenames(cls) -> list[str]:
        return ["quantize_config.json"]

    @classmethod
    def from_config(cls, config: dict[str, Any]) -> "GPTQMarlinConfig":
        dynamic = cls.get_from_keys_or(config, ["dynamic"], default={})
        dynamic = {} if dynamic is None else dynamic

        weight_bits = cls.get_from_keys(config, ["bits"])
        group_size = cls.get_from_keys(config, ["group_size"])
        desc_act = cls.get_from_keys(config, ["desc_act"])
        is_sym = cls.get_from_keys(config, ["sym"])
        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
        modules_in_block_to_quantize = cls.get_from_keys_or(
            config, ["modules_in_block_to_quantize"], default=None
        )
        return cls(
            weight_bits,
            group_size,
            desc_act,
            is_sym,
            lm_head_quantized,
            dynamic,
            config,
            modules_in_block_to_quantize,
        )

    @classmethod
    def override_quantization_method(
        cls, hf_quant_cfg, user_quant
    ) -> QuantizationMethods | None:
        can_convert = cls.is_gptq_marlin_compatible(hf_quant_cfg)

        is_valid_user_quant = (
            user_quant is None or user_quant == "marlin" or user_quant == "gptq_marlin"
        )

        if can_convert and is_valid_user_quant:
            msg = (
                "The model is convertible to {} during runtime."
                " Using {} kernel.".format(cls.get_name(), cls.get_name())
            )
            logger.info(msg)
            return cls.get_name()

        if can_convert and user_quant == "gptq":
            logger.info(
                "Detected that the model can run with gptq_marlin"
                ", however you specified quantization=gptq explicitly,"
                " so forcing gptq. Use quantization=gptq_marlin for"
                " faster inference"
            )
        return None

    def get_quant_method(
        self, layer: torch.nn.Module, prefix: str
    ) -> Optional["QuantizeMethodBase"]:
        if isinstance(layer, FusedMoE):
            from vllm.model_executor.layers.quantization.moe_wna16 import MoeWNA16Config

            if not check_moe_marlin_supports_layer(layer, self.group_size):
                logger.warning_once(
                    f"Layer '{prefix}' is not supported by GPTQMoeMarlin. "
                    "Falling back to Moe WNA16 kernels."
                )
                return MoeWNA16Config.from_config(self.full_config).get_quant_method(
                    layer, prefix
                )
            moe_quant_method = get_moe_quant_method(
                self, layer, prefix, GPTQMarlinMoEMethod
            )
            if moe_quant_method is None:
                return None
            moe_quant_method.input_dtype = get_marlin_input_dtype(prefix)
            return moe_quant_method

        quant_method = get_linear_quant_method(
            self, layer, prefix, GPTQMarlinLinearMethod
        )
        if quant_method is None:
            return None
        quant_method.input_dtype = get_marlin_input_dtype(prefix)
        return quant_method

    @classmethod
    def is_gptq_marlin_compatible(cls, quant_config: dict[str, Any]):
        quant_method = quant_config.get("quant_method", "").lower()
        num_bits = quant_config.get("bits")
        group_size = quant_config.get("group_size")
        sym = quant_config.get("sym")
        desc_act = quant_config.get("desc_act")

        if not current_platform.is_cuda():
            return False

        if quant_method != "gptq":
            return False

        # Marlin conversion is only valid if required properties are found
        if num_bits is None or group_size is None or sym is None or desc_act is None:
            return False

        if (num_bits, sym) not in cls.TYPE_MAP:
            return False

        return check_marlin_supported(
            quant_type=cls.TYPE_MAP[(num_bits, sym)], group_size=group_size
        )

    def apply_vllm_mapper(self, hf_to_vllm_mapper):
        if self.modules_in_block_to_quantize is not None:
            self.modules_in_block_to_quantize = hf_to_vllm_mapper.apply_list(
                self.modules_in_block_to_quantize
            )

    def maybe_update_config(self, model_name: str, revision: str | None = None):
        if self.modules_in_block_to_quantize:
            if is_list_of(self.modules_in_block_to_quantize, list):
                # original modules_in_block_to_quantize: list[list[str]]
                # flatten original modules_in_block_to_quantize
                self.modules_in_block_to_quantize = [
                    item
                    for sublist in self.modules_in_block_to_quantize
                    for item in sublist
                ]
            return

        unquant_dtypes = [torch.float16, torch.bfloat16, torch.float32]
        metadata = get_safetensors_params_metadata(model_name, revision=revision)
        quant_layers: set[str] = {
            param_name.rsplit(".", 1)[0]
            for param_name, info in metadata.items()
            if (dtype := info.get("dtype", None))
            and _SAFETENSORS_TO_TORCH_DTYPE[dtype] not in unquant_dtypes
        }
        self.modules_in_block_to_quantize = list(quant_layers)

class GPTQMarlinLinearMethod(LinearMethodBase):
    """Linear method for GPTQ Marlin.

    Args:
        quant_config: The GPTQ Marlin quantization config.
    """

    _kernel_backends_being_used: set[str] = set()

    def __init__(self, quant_config: GPTQMarlinConfig) -> None:
        self.quant_config = quant_config
        self.input_dtype = None
        self.quant_type = self.quant_config.quant_type

        # Verify supported on platform.
        verify_marlin_supported(
            quant_type=self.quant_config.quant_type,
            group_size=self.quant_config.group_size,
        )

    def create_weights(
        self,
        layer: torch.nn.Module,
        input_size_per_partition: int,
        output_partition_sizes: list[int],
        input_size: int,
        output_size: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ) -> None:
        output_size_per_partition = sum(output_partition_sizes)
        is_row_parallel = input_size != input_size_per_partition
        weight_loader = extra_weight_attrs.get("weight_loader")
        input_dtype = self.input_dtype

        mp_linear_kernel_config = MPLinearLayerConfig(
            full_weight_shape=(input_size, output_size),
            partition_weight_shape=(
                input_size_per_partition,
                output_size_per_partition,
            ),
            weight_type=self.quant_config.quant_type,
            act_type=params_dtype if input_dtype is None else input_dtype,
            group_size=self.quant_config.group_size,
            zero_points=False,
            has_g_idx=self.quant_config.desc_act,
        )

        kernel_type = choose_mp_linear_kernel(mp_linear_kernel_config)

        if kernel_type.__name__ not in self._kernel_backends_being_used:
            logger.info("Using %s for GPTQMarlinLinearMethod", kernel_type.__name__)
            self._kernel_backends_being_used.add(kernel_type.__name__)

        # Normalize group_size
        if self.quant_config.group_size != -1:
            group_size = self.quant_config.group_size
        else:
            group_size = input_size

        # Determine sharding
        if marlin_repeat_scales_on_all_ranks(
            self.quant_config.desc_act, self.quant_config.group_size, is_row_parallel
        ):
            # By setting scale_dim == None, weight_loader will
            # repeat the scales on each GPU in TP>1 case.
            scales_and_zp_input_dim = None
            scales_and_zp_size = input_size // group_size
        else:
            # By setting scale_dim == 0, weight_loader will
            # shard the scales in TP>1 case.
            scales_and_zp_input_dim = 0
            scales_and_zp_size = input_size_per_partition // group_size

        # Quantized weights
        qweight = PackedvLLMParameter(
            data=torch.empty(
                input_size_per_partition // self.quant_config.pack_factor,
                output_size_per_partition,
                dtype=torch.int32,
            ),
            input_dim=0,
            output_dim=1,
            packed_dim=0,
            packed_factor=self.quant_config.pack_factor,
            weight_loader=weight_loader,
        )

        # Activation order
        g_idx = RowvLLMParameter(
            data=torch.empty(
                input_size_per_partition,
                dtype=torch.int32,
            ),
            input_dim=0,
            weight_loader=weight_loader,
        )

        qzeros_args = {
            "data": torch.empty(
                scales_and_zp_size,
                output_size_per_partition // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            "weight_loader": weight_loader,
        }
        weight_scale_args = {
            "data": torch.empty(
                scales_and_zp_size,
                output_size_per_partition,
                dtype=params_dtype,
            ),
            "weight_loader": weight_loader,
        }

        if scales_and_zp_input_dim is None:
            scales = ChannelQuantScaleParameter(output_dim=1, **weight_scale_args)
            qzeros = PackedColumnParameter(
                output_dim=1,
                packed_dim=1,
                packed_factor=self.quant_config.pack_factor,
                **qzeros_args,
            )

        else:
            scales = GroupQuantScaleParameter(
                output_dim=1, input_dim=0, **weight_scale_args
            )
            qzeros = PackedvLLMParameter(
                input_dim=0,
                output_dim=1,
                packed_dim=1,
                packed_factor=self.quant_config.pack_factor,
                **qzeros_args,
            )

        layer.register_parameter("qweight", qweight)
        layer.register_parameter("g_idx", g_idx)
        layer.register_parameter("scales", scales)
        layer.register_parameter("qzeros", qzeros)

        self.kernel = kernel_type(
            mp_linear_kernel_config,
            w_q_param_name="qweight",
            w_s_param_name="scales",
            w_zp_param_name="qzeros",
            w_gidx_param_name="g_idx",
        )

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        self.kernel.process_weights_after_loading(layer)

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        return self.kernel.apply_weights(layer, x, bias)